oneDNN v3.13.5 fixes matmul correctness for Intel Xe GPUs and AVX-512 CPUs

The UXL Foundation has released oneDNN version 3.13.5, a patch update addressing several correctness bugs found in the previous 3.13.4 release. The update specifically targets deep learning primitives, focusing on matrix multiplication (matmul) accuracy across different hardware architectures.

Users running AI workloads on Intel Xe-based GPUs and x64 CPUs with AVX-512 instruction sets will see improvements in computational reliability. The release also includes performance optimizations for specific matrix-vector multiplication scenarios and fixes for memory management issues in the benchdnn benchmarking tool.

Technical improvements in oneDNN v3.13.5

• GPU Correctness: Resolved a bug affecting matmul operations that utilize multiple post-operations on Intel GPUs based on the Xe architecture.
• CPU Instruction Set Fixes: Fixed correctness errors in batched float32 (f32) matmul specifically for x64 CPUs supporting Intel AVX-512.
• Scaling and Precision: Addressed mathematical inaccuracies in matmul operations involving per-N destination scales.
• Performance and Memory: Optimized f32 matmul performance for matrix-vector multiplication on x64 CPUs and corrected excessive memory allocation in benchdnn cold-cache mode for small tensors.

Why it matters

As a core component of the oneAPI ecosystem, these fixes ensure higher numerical stability for developers deploying AI models on Intel hardware. By resolving these matmul errors, the update prevents potential model drift or inference inaccuracies in production environments using Xe GPUs or AVX-512 enabled processors.